Two Trees

Two Trees
John H. Cochrane
University of Chicago
Francis A. Longstaff
The UCLA Anderson School and NBER

Pedro Santa-Clara
The UCLA Anderson School and NBER

We solve a model with two i.i.d. Lucas trees. Although the corresponding one-tree model
produces a constant price-dividend ratio and i.i.d. returns, the two-tree model produces inter-
esting asset-pricing dynamics. Investors want to rebalance their portfolios after any change
in value. Because the size of the trees is xed, prices must adjust to offset this desire. As a
result, expected returns, excess returns, and return volatility all vary through time. Returns
display serial correlation and are predictable from price-dividend ratios. Return volatility
differs fromcash-owvolatility, and return shocks can occur without news about cash ows.
Returns that are independent over time are the standard benchmark for
theory and empirical work in asset pricing. Yet, on reection, i.i.d. returns seem
impossible with multiple positive net supply assets. If a stock or a sector rises in
value, investors will try to rebalance away from it. But we cannot all rebalance,
as the average investor must hold the market portfolio. It seems that the suc-
cessful assets expected returns must rise, or some other return moment must
change, in order to induce investors to hold more of the successful securities.
We characterize the asset price and return dynamics that result from this
market-clearing mechanism in a simple context. We solve an asset-pricing
model with two Lucas (1978) trees. Each trees dividend stream follows a
geometric Brownian motion. The representative investor has log utility and
consumes the sum of the two trees dividends. Prices adjust so that investors
are happy to consume the dividends. We obtain closed-form solutions for
prices, expected returns, volatilities, correlations, and so forth. Despite its
Graduate School of Business, University of Chicago, and NBER.
The UCLA Anderson School and NBER.

John Cochrane gratefully acknowledges research support from an NSF grant administered by the NBER and
from the CRSP. The authors are grateful for the comments and suggestions of Chris Adcock, Yacine At-Sahalia,
Andrew Ang, Ravi Bansal, Geert Bekaert, Peter Bossaerts, Michael Brandt, George Constantinides, Vito Gala,
Mark Grinblatt, Lars Peter Hansen, John Heaton, Jun Liu, Ian Martin, Anna Pavlova, Monika Piazzesi, Rene
Stulz, Raman Uppal, Pietro Veronesi, anonymous referees, and many seminar and conference participants. We
are also grateful to Bruno Miranda for research assistance.
Address correspondence to Francis Longstaff, UCLA/Anderson School, 110 Westwood Plaza, Los Angeles, CA
90095-1481; United States or email: francis.longstaff@anderson.ucla.edu.
C The Author 2007. Published by Oxford University Press on behalf of the Society for Financial Studies. All
rights reserved. For Permission, Please e-mail: journals.permissions@oxfordjournals.org
doi:10.1093/rfs/hhm059 Advance Access publication November 20, 2007
The Review of Financial Studies/ v 21 n 1 2008
simple ingredients, and although the corresponding one-tree model produces
a constant price-dividend ratio and i.i.d. returns, the two-tree model displays
interesting dynamics.
The two trees can represent industries or other characteristic-based group-
ings. The two trees can represent broad asset classes such as stocks versus
bonds, or stocks and bonds versus human capital and real estate. The two trees
can represent two countries asset markets, providing a natural benchmark for
asset market dynamics in international nance.
1
Valuation ratios and market
values of the two trees vary over time, so portfolio strategies that hold assets
based on value/growth, small/large, momentum, or related characteristics give
different average returns. Many interesting results continue to hold as one tree
becomes vanishingly small relative to the other, so the model has implications
for expected returns and dynamics of one asset relative to a much larger market.
Underlying the dynamics, we nd that expected returns typically rise with
a trees share of dividends, to attract investors to hold that larger share. As a
result, a positive dividend shock, which increases current prices and returns,
also typically raises subsequent expected returns. Thus, returns tend to display
positive autocorrelation or momentum, prices typically seem to underreact
or not to fully adjust to dividend news, and to drift upward for some time
after that news. However, there are also parameters and regions of the state
space in which expected returns decline as functions of the dividend share,
leading to mean reversion, price overreaction, and downward drift, with
corresponding excess volatility of prices and returns.
When one asset has a positive dividend shock, this shock lowers the share
of the other asset, so the expected return of the other asset typically declines.
We see negative cross-serial correlation. We see movements in the other assets
price even with no news about that assets dividends, a discount rate effect,
and another source and formof apparent excess volatility. Finally, we see that
asset returns can be positively contemporaneously correlated with each other
even when their underlying dividends are independent. The lower expected
return raises the price of the other asset. A common factor or contagion
emerges in asset returns even though there is no common factor in cash ows.
Because price-dividend ratios vary despite i.i.d. dividend growth, price-
dividend ratios forecast returns in the time series and in the cross section. Thus,
we see value and growth effects: high price-dividend ratio assets have low
expected returns and vice versa. We see that times in which a given asset has a
high price-dividend ratio are times when that asset has a low expected return.
Thinking of the two trees as stocks versus all other assets (bonds, real estate,
human capital), price-dividend ratios forecast stock index returns. These effects
coexist with the positive short-run autocorrelation of returns described above.
1
Hau and Rey (2004), Guibaud and Coeurdacier (2006), and Pavlova and Rigobon (2007) are some recent articles
that use multiple-tree frameworks to model countries.
348
Two Trees
However, although these dynamics are supercially reminiscent of those in
the empirical asset pricing literature, and although we have used the terminology
of that literature to help describe the models dynamics, we do not claimthat our
model quantitatively matches the empirical literature. Our ingredientsthe log
utility function, the pure-endowment production structure, and especially the
geometric Brownian motion dividend processesare simple but empirically
unrealistic.
Our model is simple because our goal is purely theoretical: to understand
the dynamics induced by the market-clearing mechanism. Other papers in the
emerging literature that price multiple cash-ow processes, including Bansal
et al. (2002), Menzly et al. (2004), and Santos and Veronesi (2006), include
non-i.i.d. dividend processes and temporally nonseparable preferences in order
to better t some aspects of the data. Other papers in the emerging general-
equilibrium literature such as Gomes et al. (2003) and Gala (2006) add an
interesting investment and production side to endogenize cash-ow dynamics.
For our aim, less is more. If we were to add these kinds of ingredients, we
would no longer see what dynamics result from market clearing alone versus
what dynamics come from the temporal structure of preferences, cash ows,
or technology. We also would be forced to more complex and less transparent
solution methods. As the standard one-tree model, though unrealistic, delivers
useful insights into key asset-pricing issues, this simple two-tree model can
isolate market-clearing effects that will be part of the story in more complex
models, and allows us a clearer economic intuition for those effects.
Higher risk aversion and greater numbers of trees are desirable extensions
that would not by themselves introduce dynamics. The rst might raise the
magnitudes of market-clearing dynamics, providing a better t to the data, and
the second is clearly important in its own right. However, our solution method
works only for log utility and two trees.
Was it a mistake to believe in the logical possibility if not the reality of
i.i.d. returns for all these years, given that our world does contain multiple
nonzero net supply assets? The answer is no; it is possible to construct multiple
nonzero net supply asset models with i.i.d. returns. For example, this can occur
if the supply of assets changes instantly to accommodate changes in demand.
If a price rise is instantly matched by a share repurchase, and a price decline is
instantly matched by a share issuance, then the market value of each security can
remain constant despite any variation in prices. No change in return moments
is necessary, because no change in the market portfolio occurs. In this situation,
investors can and do collectively rebalance.
In economics language, this situation is equivalent to the assumption of linear
technologies. Output is a linear function of capital with no adjustment costs,
diminishing returns, or irreversibilities, in contrast to our assumption of xed
endowments (trees, share supply). Investors can then instantly and costlessly
transfer physical capital from successful projects to unsuccessful ones or to
consumption, keeping all market weights constant. Such a linear technology
349
assumption is explicit, for example, in Cox et al. (1985). Most simply put,
we can just write the i.i.d. rate of return processes or technologies, and ask
investors how much they want to hold, allowing them collectively to hold as
much or as little as they want at any time. The resulting model will of course
have i.i.d. returns.
Although such models are possible, these clearly unrealistic supply or tech-
nological underpinnings of i.i.d. returns are perhaps not often appreciated. It is
a mistake to think that i.i.d. returns emerge from a multiperiod version of the
usual market-equilibrium derivation of the CAPM, in which demand adjusts to
a xed supply of shares. And as the last example makes clear, we should really
think of models that deliver i.i.d. returns in this way as asset quantity models,
not asset pricing models. They are models of the composition of the market
portfolio, because the asset prices and expected returns are given exogenously.
Which is the right assumption? In reality, market portfolio weights do change
over time. Thus, a realistic model should have at least some short-run adjust-
ment costs, irreversibilities, and other impediments to aggregate rebalancing.
It will therefore contain some market-clearing dynamics of the sort we isolate
and study in our simple two-tree exchange economy. On the other hand, new
investment is made, new shares are issued, and some old capital is allowed to
depreciate or is reallocated to new uses. Thus, a realistic model cannot specify
a pure endowment structure. It needs some mechanism that allows aggregate
rebalancing in the long run. The dynamics of the sort we study will apply less
and less at longer horizons. In any case, this discussion and the examples of this
paper make clear that the technological underpinnings of asset pricing mod-
els are more important to asset-price dynamics than is commonly recognized.
Dynamics do not depend on preferences alone.
1. Model and Results
1.1 Model setup
The representative investor has log utility,
U
t
= E
t
__

0
e
ln (C
t +
) d
_
. (1)
There are two trees. Each tree generates a dividend stream D
i
dt . The dividends
follow geometric Brownian motions with identical parameters,
d D
i
D
i
= dt + dZ
i
, (2)
where i = 1, 2, and dZ
i
are standard Brownian motions, uncorrelated with
each other.
To make the notation more transparent, we suppress time indices, for ex-
ample, d D
i
/D
i
d D
i t
/D
i t
, etc., unless needed for clarity. Also, we focus on
the rst asset and suppress its index, for example, d D/D = d D
1
/D
1
. Finally,
350
Two Trees
since instantaneous moments are of order dt , we will typically omit the dt term
in expressions for moments.
This is an endowment economy, so prices adjust until consumption equals
the sum of the dividends, C = D
1
+ D
2
.
This economy is a straightforward generalization of the well-known one-tree
model, and the one-tree model is the limit of our two-tree model as either tree
becomes dominant. In the one-tree model, the price-dividend ratio is a constant,
P/D = 1/, and returns are i.i.d.
1.2 Dividend share dynamics
The relative sizes of the two trees give a single state variable for this economy.
We nd it convenient to capture that state by the dividend share,
s =
D
1
D
1
+ D
2
. (3)
Expected returns and other variables are functions of state, so we can understand
their dynamics by understanding those functions of state and understanding how
the state variable s evolves.
Applying It os Lemma to the denition in Equation (3), we obtain the dy-
namics for the dividend share process,
ds = 2
2
s(1 s) (s 1/2) dt +s(1 s)(dZ
1
dZ
2
). (4)
The drift of the dividend share process is an S-shaped function. The drift is
zero when s equals 0, 1/2, or 1. The drift is positive for s between 0 and 1/2
and is negative for s between 1/2 and 1. Thus, there is a tendency for the
dividend share to mean revert toward a value of 1/2. The two dividend-growth
rates are independent, and there is no force raising an assets dividend-growth
rate if its share becomes small. Mean reversion in the share results from the
nonlinear nature of the share denition in Equation (3) through second-order
It os Lemma effectsthe drift is zero when
2
= 0. In general, however, the
drift is small so the dividend share is a highly persistent state variablethe
path of its conditional means tends only slowly back to 1/2. For example, at
s = 1/4, the drift is 3/32
2
, and with = 0.20 that implies a drift of only
0.375 percentage points per year.
Share volatility is a quadratic function of the share, largest when the trees are
of equal size, s = 1/2, and declining to zero at s = 0 or s = 1. Share volatility is
substantial. For example, at s = 1/2, and with = 0.20, share volatility is ve
percentage points per year. In turn, this means that if expected returns are func-
tions of the share, then they have the potential to vary signicantly over time.
The dispersing effects of volatility overwhelm the mean-reverting effects
of the drift, so this share process does not have a stationary distribution. A
share process with a stationary distribution might be more appealing, but that
would require putting dynamics in the dividend processes, so that small assets
catch up. We want to make it clear that all dynamics in this model come from
351
market-clearing, not from dynamics of the inputs. We return to the issue of the
nonstationary share and its effects on asset pricing below, after we see how
asset prices behave in the model.
1.3 Consumption dynamics
In the one-tree model, consumption equals the dividend, so consumption growth
is i.i.d. with constant mean E[dC/C] = and variance Var[dC/C] =
2
.
In the two-tree model, aggregate consumption C = D
1
+ D
2
follows
dC
C
= dt + s dZ
1
+ (1 s) dZ
2
. (5)
Consumption growth is no longer i.i.d. Mean consumption growth is still con-
stant, but consumption volatility is a convex quadratic function of the share,
Var
t
_
dC
C
_
=
2
_
s
2
+(1 s)
2
_
. (6)
Consumption-growth volatility is still
2
at the limits s = 0 and s = 1, but
declines to one-half that value at s = 1/2. Volatility is lower for intermediate
values of the dividend share as consumption is then diversied between the two
dividends.
1.4 The riskless rate
The investors rst-order conditions imply that marginal utility is a discount
factor that prices assets, i.e.,
M
t
=
e
t
C
t
. (7)
The instantaneous interest rate is given from the discount factor by
r dt = E
t
_
dM
t
M
t
_
= dt + E
t
_
dC
C
_
Var
t
_
dC
C
_
. (8)
In the one-tree model, consumption equals the dividend so we have
r = +
2
. (9)
We see the standard discount rate (), consumption growth (), and precaution-
ary savings (
2
) effects. Because the riskless rate is constant, the entire term
structure is constant and at. (We can compute a riskless rate, even though the
riskless asset is in zero net supply.)
Substituting the moments of the consumption dynamics, Equation (5), into
Equation (8), the interest rate in the two-tree model is
r = +
2
_
s
2
+(1 s)
2
_
. (10)
352
Two Trees
Thus, the riskless rate varies over time as a quadratic function of the dividend
share. The riskless rate is higher for intermediate values of the dividend share
because dividend diversication lowers consumption volatility, which lowers
the precautionary savings motive. Because the interest rate is not constant, the
term structure is not at.
1.5 Market portfolio price and return
As is usual in log utility models, the price-dividend ratio V
M
for the market
portfolio or claim to consumption stream is a constant
V
Mt

P
Mt
C
t
=
1
C
t
E
t
__

0
M
t +
M
t
C
t +
d
_
= E
t
__

0
e
C
t +
C
t +
d
_
=
1
. (11)
This calculation is the same for the one-tree and two-tree models, and it is valid
for all consumption dynamics.
The total instantaneous return R
M
on the market equals price appreciation
plus the dividend yield,
R
M
=
d P
M
P
M
+
C
P
M
dt =
dC
C
+ dt. (12)
[In the second equality, we use the fact from Equation (11) that P
M
= C/.]
Substituting in consumption dynamics, we nd for the one-tree model that the
market return is i.i.d.
R
M
= ( +) dt + dZ, (13)
with constant expected return and variance:
E
t
[R
M
] = ( +), (14)
Var
t
[R
M
] =
2
. (15)
For the two-tree model, the consumption dynamics in Equation (5) imply
R
M
= ( +) dt + s dZ
1
+ (1 s) dZ
2
. (16)
The expected market return and variance are now
E
t
[R
M
] = ( +), (17)
Var
t
[R
M
] =
2
[s
2
+(1 s)
2
]. (18)
The expected return on the market is the same as in the one-tree case, but the
variance of the market return equals the variance of consumption growth, which
is now a quadratic function of the dividend share, declining for intermediate
shares.
353
Subtracting the riskless rate in Equation (10) from the expected market
return in Equation (17) shows that the equity premium equals the variance of
the market return:
E
t
[R
M
] r = Var
t
[R
M
], (19)
as is usual in log utility models. From Equation (18), the variance of the market
is a convex quadratic function of the dividend share. This fact means that the
equity premium and market Sharpe ratios are also time varying, and increase
as the market becomes more polarized.
1.6 The price-dividend ratio
The price P of the rst asset is given by
P
t
= E
t
__

0
e
C
t
C
t +
D
t +
d
_
. (20)
Again, we suppress asset and time subscripts unless necessary for clarity, and
we focus on the rst asset because the second follows by symmetry.
For the remainder of this section, we impose the parameter restriction =
2
. This restriction, in conjunction with the symmetry of the assets, gives
much simpler formulas and more transparent intuition than the general case.
This restriction is not unreasonable: with = 0.20, = 0.04. In the following
section, we treat the general case that breaks this restriction, allows the trees
to have different values of and , and allows correlated shocks. We present
formulas using whichever of or
2
gives a more intuitive appearance.
Using the denition of the dividend share in Equation (3), Equation (20) can
be rewritten to give the price-consumption ratio for the rst asset:
P
C
= E
t
__

0
e
s
t +
d
_
. (21)
Valuing the asset is formally identical to risk-neutral pricing (using the discount
rate ) of an asset that pays a cash owequal to the dividend share. The dividend
share plays a similar role in many tractable models of long-lived cash ows,
including Bansal et al. (2002), Menzly et al. (2004), Longstaff and Piazzesi
(2004), and Santos and Veronesi (2006).
Solving Equation (21) with share dynamics from Equation (4), the price-
dividend ratio of the rst asset is
V
P
D
=
1
s
P
C
=
1
2s
_
1 +
_
1 s
s
_
ln(1 s)
_
s
1 s
_
ln(s)
_
. (22)
The Appendix gives a short proof, along the following lines. First, because s
follows a Markov process, Equation (21) implies that P/C is a function of s.
Using It os Lemma and the dynamics for s from Equation (4), we obtain an
354
Two Trees
expression for E[d(P/C)] in terms of the rst and second derivatives of P/C.
Second, from Equation (21), the price-consumption ratio satises
E
_
d
_
P
C
__
=
P
C
s. (23)
Equating these two expressions for E[d(P/C)], we obtain a differential equa-
tion for P/C. We then verify that Equation (22) solves this differential equation.
The formula in Equation (22) and subsequent ones are all given in terms
of elementary functions, and thus they can be characterized straightforwardly.
However, it is easier and makes for better reading simply to plot the functions,
so we followthat route in our analysis. For comparability, we use the parameter
values = 0.04, = 0.02, and = 0.20 throughout all the plots presented in
this section.
Figure 1 plots the price-consumption and price-dividend ratios for = 0.04.
The price-consumption ratio lies close to a linear function of the dividend share.
This behavior is largely a scale effect; larger dividends command higher prices.
The slightly S-shaped deviations from linearity are thus the most interesting
features of this graph. The price-consumption ratio initially is higher than the
share and then falls below the share. We anticipate that the rst asset will have
a high price at low shares and a low price at high shares.
The price-dividend ratio varies greatly from the constant value P/D =
1/ = 25 of the one-tree model. This variation of the price-dividend ratio
as a function of the dividend share drives the return dynamics that follow.
The price-dividend ratio is equal to the price-consumption ratio divided by
share s, so its behavior is driven by the small deviations from linearity in the
price-consumption ratio function.
The price-dividend ratio in Figure 1 is largest for small shares, declines until
its value is less than the price-dividend ratio of the market portfolio starting at
s = 0.5, and then increases to equal the market price-dividend ratio of 25 when
s = 1. This initially surprising nonmonotonic behavior is necessary, because
the constant price-dividend ratio of the market portfolio must equal the share-
weighted mean of the price-dividend ratios of the individual assets. If the
price-dividend ratio of the rst asset is greater than that of the market at a share
of, say, 0.25, then by symmetry, it must be less than that of the market at a share
of 0.75. It must then recover to equal the market price-dividend ratio when it
is the market at s = 1. The S-shape of the price-consumption ratio about the
45-degree line is exactly symmetric in this way, reecting the symmetry of
Equation (22).
The price-dividend ratio increases rapidly as the dividend share decreases
toward zero. The basic mechanismfor this behavior is a decline in risk premium.
As the rst asset share declines, its dividends become less correlated with
aggregate consumption. This fact lowers their risk premium and discount rate,
355
Figure 1
Price-consumption ratio and price-dividend ratio
The top panel presents the price-consumption ratio of the rst asset, in the simple case, using parameters
= 0.02, =
2
= 0.04. The bottom panel presents the price-dividend ratio. The dashed line in the top panel
gives the 45-degree line; the dashed line in the bottom panel gives the price-dividend ratio of the market portfolio
and of the one-tree model, 1/ = 25.
raising their valuations. An asset with a small share is more valuable from a
diversication perspective.
As s 0, the price-dividend ratio rises to innity. As s 0, the rst
trees dividend becomes completely uncorrelated with consumption, because
consumption consists entirely of the second trees dividend. As a result, the rst
tree is valued as a risk-free security. In this parameterization, the s = 0 limit of
the interest rate equals the dividend-growth rate r = , so the price-dividend
ratio of the growing dividend stream explodes. Although the rise in the price-
dividend ratio as s 0 is generic, because the asset becomes less risky, the
356
Two Trees
left limit is nite for other parameterizations in which the interest rate exceeds
the mean dividend-growth rate.
1.7 Asset returns and moments
Returns follow now that we know prices. Again suppressing asset and time
subscripts, the instantaneous return R of the rst asset is
R
D
P
dt +
d P
P
. (24)
For both calculations and intuition, it is convenient to express this return in
terms of the price-dividend or valuation ratio,
R =
1
V
dt +
d D
D
+
dV
V
+
d D
D
dV
V
. (25)
The return equals the dividend yield, the dividend growth, the change in valu-
ation or growth in the price-dividend ratio, and an It o term.
Although the price-dividend ratio is constant in the one-tree model, the last
two terms in Equation (25) are zero and the asset return is just the dividend
yield plus dividend (consumption) growth. With two assets, as we have seen,
the price-dividend ratio is no longer constant and varies through time as the
relative weights of the assets evolve, so the latter two terms matter. Both the
mean and variance of returns vary over time as functions of the state variable s.
Now we can nd return moments. Taking the expectation of Equation (25),
the expected return of the rst asset is
E
t
[R] =
1
V
dt + E
t
_
d D
D
_
+ E
t
_
dV
V
_
+ Cov
t
_
d D
D
,
dV
V
_
. (26)
The instantaneous variance of the rst assets return is
Var
t
[R] = Var
t
_
d D
D
+
dV
V
_
= Var
t
_
d D
D
_
+Var
t
_
dV
V
_
+2 Cov
t
_
d D
D
,
dV
V
_
. (27)
In the single-asset model with a constant valuation ratio, the variance of returns
is equal to the variance of dividend growth, the rst termin Equation (27). In the
two-asset model, variation in the price-dividend ratio can provide additional
return volatility through the second term in Equation (27). However, if the
price-dividend ratio is strongly negatively correlated with dividend growth
if an increase in dividends and thus share strongly reduces the price-dividend
ratiothen return volatility can be less than dividend-growth volatility through
the third term in Equation (27).
Applying It os Lemma to the price-dividend ratio V in Equation (22) and
using the share process in Equation (4) gives the following expressions for
357
moments of dV/V in Equations (26) and (27):
E
t
_
dV
V
_
= (1 +3(1 s))
1
V
(1 2 ln(s)) , (28)
Var
t
_
dV
V
_
=
2
_
(1 +(1 s)) +
1
1 s
ln(s)
V
_
2
, (29)
Cov
t
_
d D
D
,
dV
V
_
=
_
(1 +(1 s)) +
1
1 s
ln(s)
V
_
. (30)
Substituting the expected dividend-growth rate E
t
[d D/D] = and Equa-
tions (28) and (30) into Equation (26) and rearranging gives us the expected
return as a function of state,
E
t
[R] = +2(1 s) +
_
1
s
1 s
_
ln(s)
V
. (31)
Subtracting the riskless rate in Equation (10) from the expected return in
Equation (31) gives the expected excess return
E
t
[R] r = 2(1 s)
2
+
_
1
s
1 s
_
ln(s)
V
. (32)
Similarly, substituting from Equations (29) and (30) into Equation (27) gives
the return variance
Var
t
[R] =

2
+2
_
1
2
+(1 s) +
1
1
1 s
ln(s)
V
_
2
. (33)
Figure 2 plots the expected return given in Equation (31) as a function of the
dividend share. Expected returns rise with the share, reach an interior maximum,
and then decline slightly. This behavior of expected returns in Figure 2 mirrors
the behavior of the price-dividend ratio in Figure 1. With constant expected
dividend growth, expected returns are the only reason price-dividend ratios
vary at all, so low expected returns must correspond to high price-dividend
ratios. One way to understand the nonmonotonic behavior of expected returns,
then, is as the mirror image of the nonmonotonic behavior of the price-dividend
ratio studied above. Expected returns must be higher than the market expected
return for high shares (near s = 0.8) so that expected returns can be lower than
the market expected return for small shares.
The behavior of expected returns as a function of state in Figure 2 drives
asset return dynamics. A positive shock to the rst assets dividends increases
the dividend share. For share values below about 0.80, this event increases
expected returns. In this range, then, a positive dividend shock leads to a string
of expected price increases. Prices will seem to underreact and slowly
358
Two Trees
Figure 2
Expected return and components
E[R] gives the expected return of the rst asset as a function of the share of the rst asset. The remaining lines
give components of this expected return. 1/V gives the dividend-price ratio. E[dD/D] gives the expected dividend-
growth rate. E[dV/V] gives the expected change in the price-dividend ratio. Cov[D,V] gives Cov[dV/V,dD/D],
the covariance of dividend growth with price-dividend ratio shocks.
incorporate dividend news. To the extent that own-dividend shocks dominate
other-dividend shocks as a source of price movement, we expect to see here
positive autocorrelation and momentum of returns.
For share values above about 0.80, however, expected returns decline in the
dividend share. Here, a positive own-dividend shock leads to lower subsequent
expected returns; we see overreaction to or mean reversion after the divi-
dend shock, and we expect to see negative autocorrelation, mean reversion, and
excess volatility of returns.
Equation (26) expresses the expected return as a sum of four components.
The individual components are also plotted in Figure 2. As illustrated, the
dividend yield 1/V is generally the largest component of expected returns,
followed closely by the expected growth rate of dividends E[d D/D]. For small
dividend shares, the negative covariance between dividends and the valuation
ratio reduces the expected return substantially. The negative covariance appears
because a positive shock to dividends increases the share, and this has a strong
negative effect on the valuation ratio as per Figure 1. The expected change in
the valuation ratio is small, generally positive, and has its largest effect on the
expected return for small to intermediate values of the dividend ratio. All three
of these components change sign or slope at the point where the price-dividend
ratio starts rising as a function of share, for high share values.
359
Expected excess returns represent risk premia, reecting the covariance of
returns with consumption growth,
E
t
[R] r = Cov
t
_
R,
dC
C
_
. (34)
We can express this risk premium as the sum of a cash-ow beta, cor-
responding to the covariance of dividend growth with consumption growth,
and a valuation beta, corresponding to the covariance of valuation shocks
with consumption growth. Substituting Equation (25) into Equation (34),
we have
E
t
[R] r = Cov
t
_
d D
D
,
dC
C
_
+Cov
t
_
dV
V
,
dC
C
_
. (35)
We can evaluate the terms on the right-hand side using the dividend-growth
and consumption processes in Equations (2) and (5),
Cov
t
_
d D
D
,
dC
C
_
=
2
s = s, (36)
Cov
t
_
dV
V
,
dC
C
_
= Cov
t
_
d D
D
,
dV
V
_
(2s 1)
= (1 2s)
_
(1 +(1 s)) +
1
1 s
ln(s)
V
_
. (37)
In the last equality, we have substituted from Equation (30).
The cash-ow beta expresses what would happen if price-dividend ratios
were constant. Then, the covariance of returns with consumption growth would
be exactly proportional to the covariance of dividend growth with consumption
growth. This covariance would of course be larger precisely as the rst trees
dividend provides a larger share of consumption. Thus, the cash-ow beta is
linear in the share.
The valuation beta is more interesting, as it captures the fact that price-
dividend ratios change as well, and changes in valuation that covary with the
consumption growth generate a risk premium. Valuation betas capture the fact
that return dynamicschanges in expected return, which change valuations
spill over to the level of the expected return, as in Merton (1973).
Figure 3 plots expected excess returns from Equation (35), the risk-free rate
from Equation (10), and the cash-ow and valuation betas from Equations
(36) and (37), respectively. As shown, the expected excess return starts at zero,
but then increases rapidly as the dividend share increases. Expected excess
returns rise uniformly following a positive dividend shock, so we expect to
see the positive autocorrelation dynamics throughout the share range using this
measure (of course, expected excess returns remain positive as s 0 if one
allows correlated cash ows).
360
Two Trees
Figure 3
Expected excess return and components
E[R] r gives the expected excess return of the rst asset as a function of its share. Cov[D,C] gives the covariance
of dividend and consumption-growth shocks, Cov[dD/D,dC/C]. Cov[V,C] gives the covariance of price-dividend
ratio and consumption-growth shocks, Cov[dV/V,dC/C]. The latter two components add up to the expected excess
return. The riskless rate is given by r.
The valuation beta can have either sign. It is slightly negative for dividend
share values between about 0.50 and 0.80. For small shares, the valuation beta
is dominant. For small shares, the risk premium is due primarily to changes
in valuations correlated with the market discount rate effectrather than
changes in dividends or cash ows correlated with the market. An observer
might be puzzled why there is so much return correlation, beta, and expected
return in the face of so little correlation of cash ows.
In the limit as s 0, the expected excess return collapses to zero. (One
can show this fact analytically by taking limits of the above expressions.) In
that limit, the rst trees dividends are completely uncorrelated with consump-
tion. However, as Figure 3 demonstrates, the expected excess return rises very
quickly fromzero and is substantial even for very small shares. For example, the
expected excess return is already 1/2% at s = 0.1%. Formally, the derivative of
expected excess return with respect to share rises to innity as s 0. There-
fore, the market-clearing-induced risk premium and return dynamics remain
important for small assets.
Figure 4 plots the return volatility given in Equation (33) along with
the components of that volatility from Equation (27). Dividend-growth
Var[d D/D] gives a constant contribution of 20% volatility. Changes in val-
uation Var[dV/V] add a small amount of volatility at small share values,
where the valuation in Figure 1 is a strong function of the share. The negative
361
Figure 4
Return volatility and components
The solid line with symbols labeled Var[R] gives the variance of the rst assets return as a function of the rst
assets share. (The vertical axis labels give percent standard deviation. However, the vertical axis plots variance,
so that the variance components add up.) The horizontal solid line labeled Var[dD/D] gives the dividend-growth
component of return variance. The dashed line labeled Var[dD/D]+Var[dV/V] adds the dividend-growth variance
and the variance of the price-dividend ratio. Adding the third, usually negative, component Cov[dD/D,dV/V]
gives the overall return variance.
value of the covariance term in Equation (33) is a larger effect and pushes over-
all variance below dividend-growth variance for s < 0.80. The price-dividend
ratio is a declining function of share here, so a positive dividend shock lowers
the price-dividend ratio (Figure 1). As a result, returns (price plus dividend)
move less than dividends themselves. As the share increases, however, the
covariance term eventually becomes positive, where the price-dividend ratio
rises with the share, and adds to the total return variance. Thus, there is a small
region of excess volatility in which the volatility of the assets return exceeds
the volatility of the underlying cash ows or dividends. Again, the derivative of
volatility with respect to share becomes innite as s 0, so market-clearing
effects apply to very small assets.
Because the price-dividend ratio, expected return, and expected excess return
are all functions of the share, we can substitute out the share and plot expected
returns and excess returns as functions of the dividend yield. Figure 5 presents
the results.
Both expected returns and expected excess returns are increasing functions
of the dividend yield. Therefore, dividend yields (or price-dividend ratios)
forecast returns in the time series and in the cross section. Expected returns
show a nicely linear relation to dividend yields through most of the relevant
range. Expected excess returns showan intriguing nonlinear relation. The slope
362
Two Trees
Figure 5
Dividend yields and expected returns
The solid line plots the rst assets expected return versus its dividend yield. The dashed line plots the rst assets
expected excess return versus its dividend yield. Symbols mark the points s = 0, s = 0.1, s = 0.2, etc., starting
from the left.
of the return line is about 1.6 through the linear portion. A slope of one means
that higher dividend yields translate to higher expected returns one-for-one.
Higher slopes mean that a high dividend yield forecasts valuation increases as
well.
1.8 Market betas
With log utility, expected returns follow a conditional CAPM and consumption
CAPM. Thus, we can also understand expected excess returns by reference to
the assets beta and the market expected excess return.
In the single-tree model, the asset is the market, and its beta equals one. In the
two-tree model, the beta of each asset varies over time with the dividend share.
Using the fact from Equation (12) that the market return equals consumption
growth plus the discount rate, we have
=
Cov
t
[R, R
M
]
Var
t
[R
M
]
=
Cov
t
[R, dC/C]
Var
t
[R
M
]
. (38)
Substituting the covariances fromEquations (36) and (37), and using the market
return variance in Equation (18), we obtain
=
2
2
(1 s)
2
+1
s
1s
ln(s)
V
2
_
s
2
+(1 s)
2
_ . (39)
363
Figure 6
Market betas and market risk premium
The solid line labeled gives the beta on the market portfolio of the rst asset return. Its values are plotted on the
left vertical axis. The dashed line labeled E[R
M
] r gives the expected excess return of the market portfolio.
Its values are plotted on the right vertical axis.
(We present Equation (39) in terms of
2
, which is more intuitive for a second
moment, but this formula is only valid under the restriction
2
= of this
section.)
Figure 6 plots beta as a function of the dividend share, revealing interesting
dynamics. As shown, the beta is zero when the share is zero. As the share
increases, the beta rises quickly, in fact innitely quickly as s 0. As we can
see in Figure 3, this rise is due to the large valuation beta for small assets.
As the share rises, the beta continues to rise almost linearly. Here, the nearly
linear cash-ow beta of Figure 3 is at work: the rst asset contributes more
to the total market return and its beta begins to increase correspondingly.
At a share s = 0.5, the beta becomes greater than one, and then declines until
it becomes one again when the rst asset is the entire market. As before, we
can start to understand this nonmonotonic behavior by aggregation: the share-
weighted average beta must be one, so if the small asset has a beta less than
one, the larger asset must have a beta larger than one. As the share approaches
one, however, the beta begins to decrease and converges to one because the rst
asset becomes the market as s 1.
The expected excess return of Figure 3 is equal to the beta of Figure 6
times the market expected excess return; E
t
[R
M
] r =
2
_
s
2
+(1 s)
2
_
from Equations (18) and (19) and is also included in Figure 6. The decline
in the market expected excess return from s = 0 to s = 1/2 accounts for the
364
Two Trees
slightly lower rise in expected excess return in Figure 3 compared to the rise in
market beta in Figure 6 in this region. The rise in market expected excess return
from s = 1/2 to s = 1 offsets the decline in beta shown in Figure 6, allowing
the nearly linear rise in expected excess return shown in Figure 3.
In sum, although a conditional CAPM and consumption CAPM hold in
this model, one must make reference to time-varying expected excess returns,
expected excess market returns, and market betas in order to see the relations
predicted by the CAPM or consumption CAPM.
1.9 Return correlations
The returns of the two assets can be correlated even though their dividends are
independent. To see this fact, we can write from Equation (25),
Cov
t
[R
1
, R
2
] = Cov
t
_
d D
1
D
1
,
d D
2
D
2
_
+Cov
t
_
d D
1
D
1
,
dV
2
V
2
_
+Cov
t
_
d D
2
D
2
,
dV
1
V
1
_
+Cov
t
_
dV
1
V
1
,
dV
2
V
2
_
. (40)
Because the dividends are independent, the rst term on the right-hand side
is zero. If the price-dividend ratios V
i
for the two assets were constants, the
remaining three terms on the right-hand side would also be zero and the returns
of the two assets would be uncorrelated. However, the price-dividend ratios
vary over time and are correlated with each other and with the dividends. Thus,
the correlation of the assets returns is generally not equal to zero.
Figure 7 plots the correlation between the assets returns as a function of the
dividend share. As shown, the returns have a correlation above 25% for most
of the range of dividend shares, even though the underlying cash ows are not
correlated.
The mechanisms are straightforward. If tree two enjoys a positive divi-
dend shock, that event raises asset twos return. However, it also lowers
asset ones share. Lowering the share typically raises the price-dividend
ratioi.e., gives rise to a positive return for asset one. This story under-
lies the Cov
t
[d D
2
/D
2
, dV
1
/V
1
] component of Equation (40), graphed as
the [V
1
, D
2
] line of Figure 7. As expected it gives a positive contribution
to correlation for shares below s = 0.8, in which the price-dividend ratio in
Figure 1 is rising in the share. Adding the symmetrical contribution to corre-
lation from the effect of asset ones dividends on asset twos valuations gives
the dashed line marked [V
1
, D
2
] +[V
2
, D
1
] in Figure 7, and we see that these
two effects are most of the story for the overall correlation, marked [R
1
, R
2
] in
Figure 7.
Again, the correlations are zero in the limit as s 0 or s 1, but the deriva-
tive becomes innite at these limits so market-clearing-induced correlations are
large for vanishingly small assets.
365
Figure 7
Return correlation
The solid line with triangles labeled [R
1
, R
2
] gives the conditional correlation between the two assets returns,
given the dividend share s of the rst asset. The remaining lines give components of that correlation, using the
decomposition of Equation (40). For example, [V
1
, D
2
] gives the component of correlation corresponding to
Cov
_
dV
1
V
1
,
dD
2
D
2
_
.
1.10 Autocorrelation
How important are the return dynamics in the two-tree model? How much do
expected returns vary over time? Do market-clearing return dynamics vanish
for small assets?
As one way to answer these questions, Figure 8 presents the instantaneous
autocorrelation of returns. In discrete time, the autocorrelation is the regres-
sion coefcient of future returns on current returns, Cov[R
t +1
, R
t
]/Var[R
t
] =
Cov[E
t
(R
t +1
), R
t
]/Var[R
t
]. We compute a continuous-time conditional coun-
terpart to the second expression, Cov
t
[dER
t
, R
t
]/Var
t
[R
t
]. R
t
denotes the
instantaneous expected return. ER
t
denotes the expected return that is a func-
tion of the state as plotted in Figure 2. Thus, we can apply It os Lemma to
nd dER
t
and the covariance follows. The result expresses how much a return
shock raises subsequent expected returns.
The pattern of autocorrelation shown in Figure 8 is consistent with the
patterns of expected returns and excess returns illustrated in Figures 2 and 3.
The autocorrelation of returns is positive where the expected return in Figure 2
rises with the share, and the autocorrelation is negative where expected returns
in Figure 2 decline with share. Expected excess returns in Figure 3 rise with
share throughout, and we see a positive autocorrelation throughout. This result
is not automatic, as a positive return can also be caused by an increase in the
other assets dividend, which typically lowers the expected return. Figure 8
366
Two Trees
Figure 8
Return autocorrelation
The return line is calculated as Cov[dER
t
, R
t
]/Var[R
t
] where R
t
is the instantaneous return, and ER
t
= E
t
[R
t
]
is the expected value of the instantaneous return. It measures how much a unit instantaneous return raises the
subsequent expected return. The excess line does the same calculation for expected excess returns.
shows that this effect is dominated by the own-dividend shock, allowing a
positive autocorrelation to emerge. The magnitude of the autocorrelation is
small, reaching only one percentage point.
2
In the limit s 0, both autocorrelations vanish. However, the slope of the
autocorrelation tends to innity as s 0, so again market-clearing effects
remain important for vanishingly small assets.
2. The General Model
This section presents the general case of the two-asset model. Dividends still
follow geometric Brownian motions, but we allow different parameters,
d D
i
D
i
=
i
dt +
i
dZ
i
. (41)
The correlation between dZ
1
and dZ
2
is , not necessarily zero. We also allow
the discount rate and the volatility of dividend growth to differ,
2
= .
2
Autocorrelation is a common and intuitive measure of return predictability. However, lagged returns do not
contain all information that is available at time t . It is possible for returns to be predictablefor example,
by dividend yieldswhile not autocorrelated. The variance of expected returns, which is the variance of the
numerator of R
2
in a multivariate return-forecasting regression, is a more comprehensive measure. However, our
computation of the continuous-time counterpart to this quantity Var[dER
t
] does not differ enough from Figure 8
to warrant presentation.
367
2.1 Consumption dynamics
Applying It os Lemma to C = D
1
+ D
2
implies
dC
C
=
_
1
s +
2
(1 s)
_
dt +
1
s dZ
1
+
2
(1 s) dZ
2
. (42)
As before, consumption growth is no longer i.i.d. through time. The mean
consumption growth,
E
t
_
dC
C
_
=
1
s +
2
(1 s), (43)
is the share-weighted mean of the dividend-growth rates and so is no longer
constant. Consumption volatility,
Var
t
_
dC
C
_
=
2
1
s
2
+
2
2
(1 s)
2
+2
1
2
s(1 s), (44)
is again a convex quadratic function of the share, lower where consumption is
diversied across the two trees.
2.2 The riskless rate
Substituting the consumption moments in Equations (43) and (44) into the
expression for the riskless rate in Equation (8) gives the riskless rate in the
general two-asset model,
r = +
1
s +
2
(1 s)
2
1
s
2
2
2
(1 s)
2
2
1
2
s(1 s). (45)
The riskless rate is again a quadratic function of the dividend share. If the
means or volatilities of the dividend streams differ, it is no longer symmetric,
however.
2.3 Market price and dynamics
The market price and its dynamics are virtually the same as in the simple case.
The price-dividend ratio V
M
for the market is still 1/, and the instantaneous
return R
M
on the market equals the percent change in aggregate consumption
plus the dividend yield as in Equation (12). Thus, the expected return and
variance of the market differ only in that the moments of the consumption
process differ in the general model,
E
t
[R
M
] = +
1
s +
2
(1 s), (46)
Var
t
[R
M
] =
2
1
s
2
+
2
2
(1 s)
2
+2
1
2
s (1 s). (47)
The equity premium again equals the variance of the market as in
Equation (19).
368
Two Trees
2.4 Dividend share dynamics
An application of It os Lemma gives the dynamics of the dividend share,
3
ds = s(1 s)
_
2
s
2
1
+(1 s)
2
2
+2 (s 1/2)
1
2
_
dt
+s(1 s)(
1
dZ
1
2
dZ
2
). (48)
The drift of this dividend share process is zero when s = 0, , or 1, where
=

1
2
+
2
2
2
1
+
2
2
2
1
2
. (49)
When lies between zero and one, the drift is positive from zero to , bringing
the share up toward , and negative from to one, bringing the share down
toward . We see a more general version of the same S-shaped mean reversion
that characterizes our simple case. The diffusion coefcient in Equation (48) is
again quadratic, implying that changes in the dividend share are most volatile
when s = 1/2. As in the simple case, one tree eventually will dominate the
other so this model does not possess a stationary share distribution.
4
2.5 Asset prices
The rst assets price-dividend ratio is still derived as a discounted present
value of future shares as in Equation (21). The share process has changed to
Equation (48), however, so the form of the solution changes. The Appendix
shows that in this general case, the price-dividend ratio of the rst asset V can
be expressed as
V =
1
(1 )(1 s)
F
_
1, 1 ; 2 ;
s
s 1
_
+
1
s
F
_
1, ; 1 +;
s 1
s
_
, (50)
where
=
_
2
+2
2
, =

2
, =
+
2
,
and
=
2
2
2
/2 +
2
1
/2,
2
=
2
1
+
2
2
2
1
2
.
3
The share process is a member of the Wright-Fisher class of diffusions. These types of diffusions are often
applied in genetic theory to characterize the evolution of genes in a population of two genetic types. For example,
Karlin and Taylor (1981, ch. 15, pp. 18488) present an example in which population shares follow a diffusion
similar to Equation (48). Also see Crow and Kimura (1970) for examples and a discussion of the asymptotic
properties of these models. The cubic drift of our share process is also closely related to that of the stochastic
Ginzburg-Landau diffusion used in superconductivity physics to model phase transitions. See Kloeden and Platen
(1992) and Katsoulakis and Kho (2001).
4
This feature parallels the asymptotic properties of Wright-Fisher gene frequency models in which one of the two
gene types ultimately becomes xed in the population.
369
F(a, b; c; z) is the standard hypergeometric function (see Abramowitz and
Stegun, (1970), Chapter 15). The hypergeometric function is dened by the
power series
F(a, b; c; z) = 1 +
a b
c 1
z +
a(a +1) b(b +1)
c(c +1) 1 2
z
2
+
a(a +1)(a +2) b(b +1)(b +2)
c(c +1)(c +2) 1 2 3
z
3
+ (51)
The hypergeometric function has an integral representation, which can be used
for numerical evaluation and as an analytic continuation beyond z < 1,
F(a, b; c; z) =
(c)
(b)(c b)
_
1
0
w
b1
(1 w)
cb1
(1 wz)
a
dw, (52)
where Re(c) > Re(b) > 0. The derivative of the hypergeometric function,
needed for It os Lemma calculations, has the simple form
d
dz
F(a, b; c; z) =
ab
c
F(a +1, b +1; c +1; z). (53)
This formula can be derived by differentiating the terms of the power series in
Equation (51) (see also Gradshteyn and Ryzhik, 2000, 9.100, 9.111). Though
the hypergeometric function may be unfamiliar to many readers, it has appeared
in a number of important asset-pricing contexts, including Merton (1973),
Ingersoll (1977), Ingersoll and Ross (1992), Albanese et al. (2001), Longstaff
(2005), and many others.
2.6 Asset returns
Given the explicit price function in Equation (50) and the functional form of
its derivatives from Equation (53), the Appendix shows that the instantaneous
return on the rst asset R can be given by a direct application of It os Lemma:
R =
_
+
1
s +
2
(1 s) +(
1
2
2
+
2
s) (s)
_
dt
+
1
[s +(s)] dZ
1
2
[s 1 +(s)] dZ
2
, (54)
where
(s) =
A(s)
B(s)
,
A(s) =
1
1
_
s
1 s
_
F
_
1, 1 ; 2 ;
s
s 1
_
1
2
_
s
1 s
_
2
F
_
2, 2 ; 3 ;
s
s 1
_
+
1
1 +
_
1 s
s
_
F
_
2, 1 +; 2 +;
s 1
s
_
,
370
Two Trees
B(s) =
1
1
_
s
1 s
_
F
_
1, 1 ; 2 ,
s
s 1
_
+
1
F
_
1, ; 1 +;
s 1
s
_
.
Taking moments, we can now express how the expected return, expected
excess return, and return volatility vary with the state variable s:
E
t
[R] = +
1
s +
2
(1 s) +
_
2
2
+
2
s
_
(s), (55)
E
t
[R] r =
2
1
s
2
+
2
2
(1 s)
2
+2
1
2
s(1 s)
+
_
2
2
+
2
s
_
(s), (56)
Var
t
[R] =
2
1
[s +(s)]
2
+
2
2
[s 1 +(s)]
2
2
1
2
[s +(s)][s 1 +(s)]. (57)
2.7 Limits
In evaluating these formulas, it is useful to have exact expressions for limits as
s 0. These limits also allow us to make some general statements about the
behavior of small assets in this model. The Appendix shows that if 1, we
have
lim
s0
V = , (58)
lim
s0
(s) = , (59)
and hence,
lim
s0
E
t
[R] = +
2
+
_
2
2
_
, (60)
lim
s0
E
t
[R] r =
_
2
2
_
+
2
2
. (61)
If > 1, we instead have
lim
s0
V = ( +
2
/2)
1
, (62)
lim
s0
(s) = 1, (63)
and hence,
lim
s0
E
t
[R] = +
2
+
1
2
2
, (64)
lim
s0
E
t
[R] r =
1
2
. (65)
The limit as s 1 is the one-tree model, of course.
The risk premium (expected excess return) of the rst asset can be greater
than zero even in the limit s 0, in two ways. First of all, it can obviously
be greater than zero if the cash ows are correlated. If > 0 in Equation (65),
371
there still is a risk premium, but one that comes entirely from the covariance
of the rst assets dividend growth with the dividend growth of the second
asset, which is now the market. Second, though, the limiting risk premium can
be positive even with = 0 per Equation (61). In this case, valuation betas
are positive and lead to a positive risk premium even in the limit. However,
the price-dividend limits show us that this result holds exactly when the price-
dividend ratio, though nite for all nonzero s, tends to an innite limit as
s 0. The special case studied in the previous section has = 1. In that
case, the price-dividend ratio (just) tends to innity, the risk premium of the
rst asset approaches zero at s 0, but its derivative with respect to s is
innite.
2.8 An example
Both as an example of the general case and for its own interest, we model one
asset as a real perpetuity with
1
= 0,
1
= 0. In an economy in which the
second asset has
2
= 0.04,
2
= 0.20, and = 0.04, the top panel of Figure 9
presents the risk premium (expected excess return) of this perpetuity. In this
case, all risk premium comes from valuation betas or discount-rate changes,
because the dividend is constant. Nonetheless, there is interesting variation in
the risk premium as a function of share, and the risk premium takes on both
signs, as do bond risk premia in the data.
We can trace much of the behavior of the risk premium back to the valuation
ratio, as usual. The bottom panel of Figure 9 presents the dividend yield
the inverse of the valuation ratioalong with the risk-free rate. Betting a
perpetuity, its dividend (coupon) yield moves with the risk-free rate. However,
it does not move one-for-one, as the yield curve is not at in this model. The
region of positive risk premium corresponds to the region of rising dividend
yield or declining price-dividend ratio. All returns here are due to shocks to the
second assets dividends. If that dividend increases, the share of the rst asset
decreases, raising the price and hence return of the rst asset. Hence, the rst
asset return is positively correlated with consumption growth, and generates
a positive risk premium. The converse logic holds in the region s > 1/2 with
a rising price-dividend ratio, declining dividend-price ratio, and negative risk
premium.
3. Concluding Remarks
We extend the classic single-asset Lucas-tree pure-exchange framework to the
case of two assets and solve the model in closed form. Our two-tree model has
the simplest ingredients, log utility, and i.i.d. normal dividend growth. Nonethe-
less, market-clearing logic and a xed-share supply generate interesting and
complex patterns of time-varying asset prices, expected returns, risk premia,
variances, covariances, and correlations.
372
Two Trees
Figure 9
Expected excess returns and dividend yield for a perpetuity
The rst asset is a perpetuity that pays a constant dividend stream. Its parameters are
1
= 0,
1
= 0. The second,
risky, asset has parameters
2
= 0.04,
2
= 0.20. The discount factor is = 0.04. The top panel displays the
expected excess return of the perpetuity as a function of its dividend share. The bottom panel plots the dividend-
price ratio of the perpetuity, along with the instantaneous interest rate.
3.1 Summary and intuition
With the results in hand, we can restate the intuition more clearly, emphasizing
the quantitatively important channels. Start at the left-hand side of the plots,
for small dividend shares. As the dividend share of the rst tree increases from
zero, the rst tree becomes a larger part of the total, so its beta and risk premium
naturally increase. Its expected excess return therefore rises fromzero as shown
in Figure 3. Also, as the rst tree becomes a larger part of the total, that total
becomes less risky by diversifying across the two trees. This change raises
373
interest rates by lowering the precautionary savings motive, again shown in
Figure 3. The expected return is the sum of the expected excess return and
the interest rate, and therefore rises even more steeply with share as shown in
Figure 2. With a constant dividend-growth rate, higher expected returns mean
lower price-dividend ratios, which is why the price-dividend ratio falls with the
share in Figure 1.
The rise of the expected return with share and the decline of the price-
dividend ratio with share underlie many of the dynamics we nd. They also
express the underlying market-clearing intuition. If there is a shock to the
rst trees dividend, investors want to rebalance. Equivalently, they want to
spread some of their larger wealth across both trees. They cannot collectively
rebalance, so prices and expected returns must adjust. The expected return
of the rst tree must rise, making it more attractive to hold the larger share,
and thus its valuation must fall. The expected return of the second tree falls
(or, if you wish, the expected return of the rst tree when there is a shock
to the second tree) and its price rises. Investors want to buy more of the
second tree but cannot, forcing its price to rise. Equivalently, the expected
return of the second tree must fall so investors will hold it in its now smaller
proportions.
However, overall betas that drive risk premia, as shown in Figure 6, depend
not only on this cash-owbeta intuition, shown as Cov[D, C] in Figure 3, but
also on valuation betas, the tendency of the price-dividend ratio to rise or fall
when the market and total consumption rise or fall. When the rst trees share
is small, most increases in aggregate consumption come from increases in the
second trees dividend. Such an increase decreases the rst trees share, which
increases the rst trees price-dividend ratio (Figure 1). Therefore, valuation
betas are positive and large in the region of small shares, and they are a
large component of risk premia. We see this in the large increase of expected
excess returns in the left-hand side of Figure 3, and its decomposition into
compensation for cash-ow risk labeled Cov[D, C] and valuation risk labeled
Cov[V, C]. The same nonlinearity is apparent in the beta of Figure 6. Figure 9
treats the case that the rst tree is a perpetuity with a constant dividend stream.
Here, the entire risk premiumshown in the top panel derives fromthis valuation-
beta mechanism.
Now let us move to the middle and right-hand parts of the graphs, where the
rst tree becomes a larger and larger share of the total. When the share reaches
one, the rst tree is the market, so it must have the market price-dividend
ratio, as shown in Figure 1. However, there must be a region as shown in the
right half of Figure 1 in which the rst trees price-dividend ratio is less than
that of the market, because here the second trees price-dividend ratio, with a
small share, is greater than that of the market. Thus, the price-dividend ratio
is not monotonic with the share, as shown in Figure 1. In terms of risk premia,
this behavior is driven by interest rates (Figure 3) for our canonical example.
The expected excess return is driven almost entirely by cash-ow betas at this
374
Two Trees
point, but interest rates fall with the share because the market becomes less
diversied and more volatile as the share rises past 1/2. The lower interest rates
eventually overcome the higher expected excess returns and cause a slight rise
in the price-dividend ratio for shares above 0.9 (Figure 1) and a slight decline
in the expected return (Figure 2).
The valuation beta is not always positiveit is possible that a rise in
the market dividend lowers the price-dividend ratio of the rst tree. Though
quantitatively small, this effect can be seen in Figure 3 for shares between
about 0.5 and 0.8. In this region, the rst tree is more than half the mar-
ket, so a rise in the market typically means a rise in the rst trees dividend
and its share. The price-dividend ratio is still downward sloping in this re-
gion (Figure 1), so a rise in the share means a decline in the price-dividend
ratio.
The possibility of a negative valuation beta is quantitatively more in-
teresting in the bond-stock case of Figure 9. Here, all movements in total
consumption come from movements in the second dividend. When the share is
above one half, the dividend yield shown in the bottom of Figure 9 declines, so
the U-shaped price-dividend ratio of the rst tree rises. Thus, an increase in the
market, which lowers the rst trees share, will also lower the rst trees price
a negative valuation beta. This negative beta generates the negative expected
excess return shown in the right half of the top panel of Figure 9. Intuitively,
when the second (small) trees share rises under this circumstance it is still
true that agents want to spread their increase in wealth across both trees, which
should raise the price of the rst tree. However, the interest rate also changes,
because both mean and variance of consumption growth have changed, and this
change more than offsets the rebalancing desire.
The remaining graphs draw out the dynamic implications of these effects.
Because expected returns (Figure 2) and excess returns (Figure 3) vary with
the share, and because the dividend yield (Figure 1) varies with the share,
dividend yields forecast stock returns and excess returns as shown in Figure
5. Returns are cross-correlated despite no correlation in dividend growth, as
shown in Figure 7. When the rst trees share is small, an increase in the
second trees dividend lowers the rst trees share, which (Figure 1) substan-
tially raises its price-dividend ratio and thus gives a shock to the rst trees
return, shown in the [V
1
, D
2
] line of Figure 7. The symmetric effect gen-
erates a positive correlation when the rst asset has a large share. Returns
are correlated over time as shown in Figure 8. An increase in dividends in-
creases todays return. This also increases the share, which increases expected
returns (Figure 2) and excess returns (Figure 3), leading to positive autocor-
relation. The small region in which expected returns (Figure 2) decline with
share, driven by the decline in interest rates with share (Figure 3), generates a
small, but theoretically interesting, region of negative return autocorrelation in
Figure 8.
375
3.2 Discussion
A natural question is: Are the effects generated by the model quantitatively
important and empirically relevant, and if not, what conclusions should one
come to?
Our rst answer is that this is the wrong question. Our aimis theory, to under-
stand and characterize the dynamics induced by the market-clearing mechanism
with the simplest possible preference and technology structure. Our aim is not
to provide a calibrated model that replicates the full range of asset-pricing facts
and puzzles. If the predictions of this model are counter to fact, one cannot con-
clude that we should ignore market-clearing dynamics. As a matter of logic,
market-clearing dynamics will be present in any model that does not allow
instantaneous aggregate rebalancing. If the models predictions are false, one
can only conclude that other ingredients are present in the real-world economy,
which is obviously true.
That said, however, the realism or unrealism of the models predictions
should be addressed to some extent. If the model gives unrealistic predictions,
it is worth speculating whether more general versions of market-clearing dy-
namics might address them, including more trees and higher risk aversion, or
at what point preferences or technologies with dynamic elements will have to
be part of the story.
The magnitudes are small. The relation between price-dividend ratios and
subsequent returns in our model is about the same as that found in empirical
time-series or cross-sectional (value-growth) return forecasts, but other effects
such as the autocorrelation of returns are smaller in our model than many claims
in the empirical literature.
However, we only use log utility, because we are not able to solve the model
for higher risk aversion. It is natural to speculate that risk premia will be larger
with higher risk aversion. Similarly, with log utility, our model is obviously
inconsistent with the observed equity premium and low consumption volatility,
but all the familiar equity-premium-boosting ingredients are likely to change
that with two trees, as they do with one tree.
More seriously than magnitudes, some of the patterns in our model seem at
odds with the data. For example, small rms are also growth rms in the
simple parameterization of our model, with high valuations and low expected
returns. Clearly, a small rm effect requires some other mechanism, such as
a larger covariance of the underlying cash ows with the aggregate.
More generally, our model has only one state variable, the dividend share,
which is both the only aggregate state variable and the only variable describing
cross-sectional variation. Our model also has only two shocks. Taken literally to
the data, it is as easy to reject these predictions as it is to reject the prediction
that there are only two assets.
One deep question is whether market-clearing effects apply to individual
assets, i.e., for very small share values, or whether they only apply to large
376
Two Trees
aggregates that are substantial shares of aggregate wealth. If the latter conclu-
sion holds, then market-clearing dynamics may not be an important part of the
story for many empirical ndings that are concentrated in the very smallest of
rms. One must recognize, however, that the question is not well posed. How
can aggregates show effects that are not present in the individual constituents
of those aggregates? If they matter for aggregates, ipso facto, in some sense
they must matter for individual rms that compose those aggregates. Also,
much empirical work on individual rm behavior in fact studies the behavior
of portfolios of such rms, which constitute an aggregate with non-negligible
share. Thus, the question is not easy to answer in the abstract. A full answer
must await the analysis of an N-tree model, and must be specic about which
fact one wants to address.
We address some of this issue in our study of the limits as s 0. In our
simple model, though expected returns, risk premia, and autocorrelations do
go to zero as s 0, the derivatives go to innity at that point for many sets
of parameters, so vanishingly small assets can have substantial risk premia and
return dynamics induced by market clearing.
The dividend process and consequent share process in our model are obvi-
ously chosen for transparency, not realism. First of all, as in all endowment
economies, it is unrealistic that there is no investment or share issue and re-
purchase at all. Our model is clearly best applied to thinking about dynamics
that occur in the short run before share issuance/repurchase or investment and
disinvestment can take place. For this reason, we do not think it too trou-
bling that plots of price-dividend ratios or average returns versus shares in
the data do not look like those of our model.
5
It seems likely that a model
with, say, adjustment costs that slow down investment and disinvestment will
generate market-clearing dynamics in the short run, meaning that changes in
shares are associated with changes in valuations and average returns, while
reverting in the long run to an i.i.d. economy in which there is no associa-
tion between the level of the share and the level of valuations and average
returns.
Finally, it may seem a little unsettling that the share in our model does
not have a stationary distribution. This result is an inescapable implication
of the geometric Brownian motion for dividends: one of two random walks
will eventually end up dominating the other one. It is not obvious what the
right assumption is here. In the end the car industry did dominate the
horse buggy industry, so perhaps birth of new industries rather than mean
reversion of dividends is the right way to generate long-run nondegenerate
shares.
The question for us is: To what extent does the fact that the distribution
of shares tends to two points (zero and one) have on the short-run asset pric-
ing dynamics we have characterized, such as autocorrelation, predictability,
5
We thank Ravi Bansal for pointing this out.
377
price-underreaction, and so forth? Is this model a good parable for what would
happen with, say, a very long-run mean reversion in the shares? The basic asset
pricing, Equation (21), reproduced here,
P
C
= E
t
__

0
e
s
t +
d
_
, (66)
gives some comfort on this point. What matters for asset pricing is the condi-
tional mean of the share, which always tends to 1/2. The fact that the distribution
underlying that mean tends to two points, 0 and 1, does not have any effect
on asset prices. We can also document that the share tends to its endpoints
very slowly. For example, when the initial dividend share is 0.50, there is only
about a 15% chance of the dividend share being below 0.05 or being above
0.95 after 100 years. Even when the initial share is 0.05, there is only about
a 52% chance of the share being below 0.05 after 100 years, and less than a
1% chance of being above 0.95 after 100 years. Because the present value of
cash ows beyond 100 years has only a negligible effect on asset values, these
simulations indicate clearly that the asymptotic nonstationarity property of the
dividend share is not driving the results. We have also repeated the analysis
by simulation with cash ows truncated at 100 years, and nd no differences
worth reporting. As a result, we suspect that our core analysis would be very
little affected if one changed to a dividend process with long-run mean rever-
sion, though of course until the alternative model is solved one cannot say for
sure.
Appendix A
A1 Derivation of asset prices, simplied case
Here we prove Equation (22) for the price-dividend ratio of the rst asset in the simplied ( =
2
)
two-tree economy. From Equation (21), the price-consumption ratio Y for the rst asset is
Y
t

P
t
C
t
= E
t
__

0
e
s
t +
d
_
. (A1)
This ratio solves
E [dY] = (Y s) dt. (A2)
From Equation (A1), Y depends only on the current value and future distribution of s. Thus,
because s follows a Markov process (from Equation (4)), Y must be a function Y(s) of the state
variable. Using It os Lemma,
E [dY(s)] = Y
(s) E [ds] +
1
2
Y
(s) E
_
ds
2
_
. (A3)
Putting together the last two equations, Y solves the differential equation
Y
(s) E [ds] +
1
2
Y
(s) E[ds
2
] = (Y s) dt. (A4)
378
Two Trees
Now we add the share process,
ds = s(1 s)(1 2s)
2
dt +s(1 s)(dZ
1
dZ
2
). (A5)
Specializing to =
2
, the differential equation becomes
s(1 s)(1 2s)Y
(s) +s
2
(1 s)
2
Y
(s) = Y(s)
s
. (A6)
We conjecture a solution and take derivatives:
Y(s) =
1
2
_
1 +
1 s
s
ln(1 s)
s
1 s
ln(s)
_
, (A7)
Y
(s) =
1
2
1
s (1 s)
_
1 +
1 s
s
ln (1 s) +
s
1 s
ln (s)
_
, (A8)
Y
(s) =
1
1
s
2
(1 s)
2
_
(2s 1)
(1 s)
2
s
ln (1 s) +
s
2
1 s
ln (s)
_
.
(A9)
Substituting into the differential equation and multiplying by 2,
0 = (1 2s)
_
1 +
1 s
s
ln (1 s) +
s
1 s
ln (s)
_
2
_
(2s 1)
(1 s)
2
s
ln (1 s) +
s
2
1 s
ln (s)
_
_
1 +
1 s
s
ln(1 s)
s
1 s
ln(s)
_
+2s. (A10)
Grouping terms,
0 = (1 2s) 2 (2s 1) 1 +2s
_
(1 2s)
1 s
s
2
(1 s)
2
s
+
1 s
s
_
ln(1 s)
_
(1 2s)
s
1 s
+2
s
2
1 s

s
1 s
_
ln(s). (A11)
Each term is zero, verifying the conjectured solution. Dividing the price-consumption ratio by s
gives the price-dividend ratio in Equation (22).
A2 Derivation of asset prices, general case
The price-consumption ratio of the rst asset is given by
P
C
= E
t
__

0
e
D
t +
C
t +
d
_
= E
t
_

0
e
1
1 +
D
2,t +
D
1,t +
d
= E
t
__

0
e
1
1 +qe
u
d
_
, (A12)
379
where q is the initial dividend ratio D
2,t
/D
1,t
and u is a normally distributed random variate with
mean and variance
2
, and where
=
2
2
2
/2 +
2
1
/2,
2
=
2
1
+
2
2
2
1
2
.
Note that dt = E[ln(D
2
/D
1
)] and
2
dt = Var[ln(D
2
/D
1
)]. Introducing the density for u into
the last integral gives
P
C
=
_

0
_

1
_
2
2
1
1 +qe
u
exp
_
(u )
2
2
2
_
du d. (A13)
Interchanging the order of integration and collecting terms in gives
P
C
=
_

1
_
2
2
1
1 +qe
u
exp
_
u
2
_
_

0
1/2
exp
_
u
2
2
2
1

2
+2
2
2
2

_
d du. (A14)
From Equation (3.471.9) of Gradshteyn and Ryzhik (2000), this expression becomes
P
C
=
_

2
_
2
2
1
1 +qe
u
exp
_
u
2
__
u
2
2
+2
2
_
1/4
K
1/2
2
_
u
2
(
2
+2
2
)
4
4
du, (A15)
where K
1/2
() is the modied Bessel function of order 1/2 (see Abramowitz and Stegun, 1970, ch.
9). From the identity relations for Bessel functions of order equal to an integer plus one half given
in Gradshteyn and Ryzhik Equation (8.469.3), however, the above expression can be expressed as
P
C
=
1
1
1 +qe
u
exp
_
u
2
_
exp
_
2
|u|
_
du, (A16)
where
=
_
2
+2
2
.
In turn, Equation (A16) can be written as
P
C
=
1
_

0
1
1 +qe
u
exp (u) du +
1
_
0
1
1 +qe
u
exp (u) du, (A17)
where
=

2
,
=
+
2
.
380
Two Trees
Dene w = e
u
. By a change of variables, Equation (A17) can be written as
P
C
=
1
q
_
1
0
1
1 +w/q
w
dw +
1
_
1
0
1
1 +qw
w
1
dw. (A18)
From Abramowitz and Stegun Equation (15.3.1), this expression becomes
P
C
=
1
q(1 )
F(1, 1 ; 2 ; 1/q) +
1
F(1, ; 1 +; q).
(A19)
Finally, substituting q = (1 s)/s into Equation (A19) and dividing by s gives the price-dividend
ratio of the rst asset,
V =
1
(1 )(1 s)
F
_
1, 1 , 2 ;
s
s 1
_
+
1
s
F
_
1, ; 1 +;
s 1
s
_
, (A20)
which is Equation (50).
The special case solution for V given in Equation (22) can also be obtained directly from the
general solution above. To see this, note that the parameter restrictions in the special case imply
that = 1 and = 1. Substituting these values into Equation (A20) results in hypergeometric
functions of the form F(1, 2; 3; ) and F(1, 1; 2; ). A repeated application of the relations for
contiguous hypergeometric functions described in Abramowitz and Stegun (1970) to F(1, 2; 3; )
allows V to be expressed entirely in terms of F(1, 1; 2; ). From Abramowitz and Stegun Equation
(15.1.3), however, F(1, 1; 2; z) = ln(1 z)/z. Substituting this into the expression for V leads
immediately to Equation (22).
By symmetry, the price-dividend ratio of the second asset is given by
V
2
=
1
(1 +)s
F
_
1, 1 +; 2 +;
s 1
s
_
1
(1 s)
F
_
1, ; 1 ;
s
s 1
_
. (A21)
Applying the recurrence relations for contiguous hypergeometric functions presented in
Abramowitz and Stegun (1970), Equations (15.2.18) and (15.2.20) give the result
P
1
+ P
2
=
C
=
D
1
+ D
2
= P
M
. (A22)
To solve for the returns on the rst asset, it is convenient to dene Y = P/C and to dene
an alternative state variable x = D
1
/D
2
. Note that s = (1 + x)/x. An application of It os Lemma
gives
dx =
_
2
+
2
2
2
_
x dt +
1
x dZ
1
2
x dZ
2
. (A23)
Now applying It os Lemma to Y gives
dY =
__
2
+
2
2
2
_
x Y
x
+
_
2
1
+
2
2
2
1
2
_
x
2
Y
xx
/2)
_
dt
+
1
x Y
x
dZ
1
2
x Y
x
dZ
2
. (A24)
381
Because P = CY, the above dynamics imply
d P
P
=
_
1
s +
2
(1 s) +
_
2
+
_
2
1
+
2
2
2
1
2
_
s
_
xY
x
/Y
+
_
2
1
+
2
2
2
1
2
_
x
2
Y
xx
/(2Y)
_
dt +
1
(xY
x
/Y +s) dZ
1
2
(xY
x
/Y +s 1) dZ
2
. (A25)
An argument similar to that in Section A1 of the Appendix can be used to show that Y satises the
differential equation
_
2
1
+
2
2
2
1
2
_
x
2
Y
xx
/2 =
_
2
+
2
2
_
xY
x
+Y x/(1 + x). (A26)
Substituting out for Y
xx
using the above expression allows us to rewrite
Equation (A25) as
d P
P
=
_
+
1
s +
2
(1 s) +
_
2
2
+
_
2
1
+
2
2
2
1
2
_
s
_
xY
x
/Y
_
dt +
1
(xY
x
/Y +s) dZ
1
2
(xY
x
/Y +s 1) dZ
2
. (A27)
Applying the expression for the derivative of the hypergeometric function repeatedly to the equation
for Y given in Equation (A19), and then substituting out for x using x = s/(1 s), shows that
xY
X
/Y is simply (s) as given in Equation (54). Substituting (s) into Equation (A27) and using
the denition for
2
gives the result.
Finally, while many numerical software programs calculate the hypergeometric function, it
is very simple to evaluate the function by numerically integrating the integral representation in
Equation (52). In doing so, however, it is important to use a numerical integration algorithm
that provides robust results for functions with endpoint singularities. As one example of such an
algorithm, see Piessens et al. (1983).
A3 Limits
In this section, we derive limits for price-dividend ratios as s 0 and s 1. Also, we derive
limits for the function (s) that gures prominently in the asset-price dynamics in Equation (54).
We focus on the rst asset, as the second is symmetric.
From the power series expression for the hypergeometric function, F(a, b; c; 0) = 1. Because
of this result, it is useful to apply the linear transformation formula given in Abramowitz and
Stegun Equation (15.3.7) so that the argument of the hypergeometric function goes to zero at the
limit being evaluated:
F(a, b; c; z) =
(c)(b a)
(b)(c a)
(z)
a
F(a, 1 c +a; 1 b +a; 1/z)
+
(c)(a b)
(a)(c b)
(z)
b
F(b, 1 c +b; 1 a +b; 1/z).
(A28)
382
Two Trees
To obtain the limit of the price-dividend ratio V as s 0, we use the linear transformation
formula to rewrite Equation (A20) as
V =
1
(1 )
_
1
1 s
_
F
_
1, 1 , 2 ;
s
s 1
_
+
1
_

1
_ _
1
1 s
_
F
_
1, 1 ; 2 ;
s
s 1
_
+
1
_
1
s
_
( +1) (1 )
_
s
1 s
_
F
_
, 0; ;
s
s 1
_
.
(A29)
From this expression, it is readily seen that
lim
s0
V =
, if 1;
1
+
2
/2
, if > 1.
(A30)
To obtain the limit of the price-dividend ratio as s 1, we again use the linear transformation
formula and rewrite Equation (A20) as
V =
1
(1 )
_
1
_ _
1
s
_
F
_
1, ; 1 ;
s 1
s
_
+
1
(1 )
_
1
1 s
_
(2 ) ()
_
1 s
s
_
1
F
_
1 , 0; 1 ;
s 1
s
_
+
1
_
1
s
_
F
_
1, ; 1 +;
s 1
s
_
. (A31)
From this, it immediately follows that
lim
s1
V =
1
. (A32)
A similar approach can be used to show that
lim
s0
V
2
=
1
, (A33)
and that
lim
s1
V
2
=
2
/2
, if < 1;
, if 1.
(A34)
Finally, the use of lHopitals rule and the repeated application of the linear transformation
formula gives
lim
s0
(s) =
_
, if 1;
1, if > 1,
(A35)
383
and
lim
s1
(s) = 0. (A36)
Substituting the limiting values of (s) into the asset-price dynamics in
Equation (54) allows us to fully characterize the properties of these price dynamics as s 0
and s 1.
References
Abramowitz, M., and I. A. Stegun. 1970. Handbook of Mathematical Functions. New York: Dover Publications.
Albanese, C., G. Campolieti, P. Carr, and A. Lipton. 2001. Black Scholes Goes Hypergeometric. Risk 14:99103.
Bansal, R., R. F. Dittman, and C. T. Lundblad. 2002. Consumption, Dividends, and the Cross Section of Stock
Returns. Working Paper, Duke University.
Cox, J., J. E. Ingersoll, and S. A. Ross. 1985. An Intertemporal General Equilibrium Model of Asset Prices.
Econometrica 53:36384.
Crow, J. F., and M. Kimura. 1970. An Introduction to Population Genetics Theory. Minneapolis: Burgess.
Gala, V. D. 2006. Investment and Returns. Working Paper, Graduate School of Business, University of Chicago.
Gomes, J. F., L. Kogan, and L. Zhang. 2003. Equilibrium Cross-Section of Returns. Journal of Political Economy
111:693732.
Gradshteyn, I. S., and I. M. Ryzhik. 2000. Tables of Integrals, Series, and Products, 6th ed. New York: Academic
Press.
Guibaud, S., and N. Coeurdacier. 2006. A Dynamic Equilibrium Model of Imperfectly Integrated Financial
Markets. Working Paper, Princeton University.
Hau, H., and H. Rey. 2004. Can Portfolio Rebalancing Explain the Dynamics of Equity Returns, Equity Flows,
and Exchange Rates? American Economic Review 94:12633.
Ingersoll, J. E. 1977. An Examination of Corporate Call Policies on Convertible Securities. Journal of Finance
32:46378.
Ingersoll, J. E., and S. A. Ross. 1992. Waiting to Invest: Investment and Uncertainty. Journal of Business
65:129.
Karlin, S., and H. M. Taylor. 1981. A Second Course in Stochastic Processes. London: Academic Press.
Katsoulakis, M. A., and A. T. Kho. 2001. Stochastic Curvature Flows: Asymptotic Derivation, Level Set, and
Formulation and Numerical Experiments. Interfaces and Free Boundaries 3:26590.
Kloeden, P. E., and E. Platen. 1992. Numerical Solution of Stochastic Differential Equations. Berlin: Springer-
Verlag.
Longstaff, F. A. 2005. Asset Pricing in Markets with Illiquid Assets. Working Paper, University of California,
Los Angeles.
Longstaff, F. A., and M. Piazzesi. 2004. Corporate Earnings and the Equity Premium. Journal of Financial
Economics 74:40121.
Lucas, R. E., Jr. 1978. Asset Prices in an Exchange Economy. Econometrica 46:142945.
Menzly, L., T. Santos, and P. Veronesi. 2004. Understanding Predictability. Journal of Political Economy 112:1
47.
Merton, R. C. 1973. An Intertemporal Capital Asset Pricing Model. Econometrica 41:86788.
384
Two Trees
Pavlova, A., and R. Rigobon. 2007. Asset Prices and Exchange Rates. Review of Financial Studies 20:113981.
Piessens, R., E. deDoncker-Kapenga, C. W. Uberhuber, and D. K. Kahaner. 1983. QUADPACK. New York:
Springer-Verlag.
Santos, T., and P. Veronesi. 2006. Labor Income and Predictable Stock Returns. Review of Financial Studies
19:144.
385

Two Trees

Caricato da

Informazioni sul documento

Descrizione originale:

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Two Trees

Caricato da

Copyright:

Formati disponibili

Two Trees

The UCLA Anderson School and NBER

The UCLA Anderson School and NBER

Graduate School of Business, University of Chicago, and NBER.

The UCLA Anderson School and NBER.

Potrebbero piacerti anche